design and experimental evaluation of multi-user beamforming in wireless lans theodoros salonidis...
TRANSCRIPT
Slide 1
Design and Experimental Evaluation of Multi-User Beamforming in Wireless LANsTheodoros SalonidisTechnicolorACM MobiCom 2010Edward KnightlyRiceNarendra AnandRiceEhsan AryafarRiceMIMO LANsEhsan AryafarRice Networks GroupEhsan AryafarRice Networks GroupRxRxRxRxRxMIMO increases throughput with antenna arrays at transmitter and receiverHowever, real world client devices have fewer antennas than APs due to cost and space MUBF allows for APs to leverage antennas belonging to group of nodesTxRxTxWe present the design and experimental evaluation of the first MUBF platform for WLANs
Xirrus 16 ant AP
Crash Course on BeamformingOmniFixed vs ant selectionEhsan AryafarRice Networks Group
p1p2APLets start on a quick background of how you can serve multiple people. In this figure we have an AP with two antennas and two clients each with one antennas.[animate] The traditional way of communicating with these users, is to serve one user at a time in a TDMA fasion.There are many different ways for the access point to use its multiple antennas. One of the simplest methods currently used by many Aps is use of a fixed antenna (or rando selection) by AP and transmit in an omni-directional way. Figure at the left shows the radiation pattern for this method. In the basic approach no feedback from clients is received.Off course, Aps could benefit from the available antenna diversity. There are several papers that have shown you can obtain huge gains by selecting the appropriate antenn. An alternative way for AP to benefit from its antennas is through switched beam. Here, the AP has an antenn array. The signal that is fed to each antenna is shifted with specific phases, such that the resulting beam has high directionality in a specific direction. The AP can then switch between a set of pre-defined beams. Compared to the Omni transmission, this scheme allows for higher received SNR at the receiver and also higher coverage.A user can further increase its received SNR by sharing its channel information to the AP and through adaptive beamforming scheme. If AP has a users channel information, it can construct its beam such that it points towards that specific user. This scheme can further increase received SNR and coverage with the additional feedback of channel information for each antenna.
Crash Course on BeamformingOmniFixed vs ant selectionEhsan AryafarRice Networks Group
p1p2APAdaptive Beam (SUBF)Higher coverageHigher SNRLets start on a quick background of how you can serve multiple people. In this figure we have an AP with two antennas and two clients each with one antennas.[animate] The traditional way of communicating with these users, is to serve one user at a time in a TDMA fasion.There are many different ways for the access point to use its multiple antennas. One of the simplest methods currently used by many Aps is use of a fixed antenna (or rando selection) by AP and transmit in an omni-directional way. Figure at the left shows the radiation pattern for this method. In the basic approach no feedback from clients is received.Off course, Aps could benefit from the available antenna diversity. There are several papers that have shown you can obtain huge gains by selecting the appropriate antenn. An alternative way for AP to benefit from its antennas is through switched beam. Here, the AP has an antenn array. The signal that is fed to each antenna is shifted with specific phases, such that the resulting beam has high directionality in a specific direction. The AP can then switch between a set of pre-defined beams. Compared to the Omni transmission, this scheme allows for higher received SNR at the receiver and also higher coverage.A user can further increase its received SNR by sharing its channel information to the AP and through adaptive beamforming scheme. If AP has a users channel information, it can construct its beam such that it points towards that specific user. This scheme can further increase received SNR and coverage with the additional feedback of channel information for each antenna.
Multi-User Beamforming: Throughput Increase Ehsan AryafarRice Networks Group
s1s2APMUBF sends the contents to both receivers at the same timeEach users data stream is weighted at the transmitter
++
desired signalinter-user interferenceAppropriate weights can reduce or eliminate the amount of inter-user interference
The best way for throughput maximization, however is to transmit to multiple receivers simult. This is accomplished by transmitting to both receivers at the same time and frequency channel, however by limiting the amount of inter-user interference.There are multiple ways to serve the users concurrently. In MUBF, each users data stream is multiplied by an appropriate weight vector and the sum of all streams is transmitted through the AP. For a user k, the received signal is .The weigth vectors are selected according to each users specific channel information and can be selected to reduce or eliminate inter-user interference.The top-right figure shows an example of two receivers with zero inter-user interference.Apart from weights, the received signal also depends on the power allocated to each data stream Ehsan AryafarRice Networks Group
s1s2AP
+
desired signalinter-user interferenceZero-Forcing beamforming (ZFBF)weights are selected such that the amount of inter-user interference is zero
Multi-User Beamforming: Throughput Increase 0The best way for throughput maximization, however is to transmit to multiple receivers simult. This is accomplished by transmitting to both receivers at the same time and frequency channel, however by limiting the amount of inter-user interference.There are multiple ways to serve the users concurrently. In MUBF, each users data stream is multiplied by an appropriate weight vector and the sum of all streams is transmitted through the AP. For a user k, the received signal is .The weigth vectors are selected according to each users specific channel information and can be selected to reduce or eliminate inter-user interference.The top-right figure shows an example of two receivers with zero inter-user interference.Apart from weights, the received signal also depends on the power allocated to each data stream Multi-User Beamforming: Interference ReductionEhsan AryafarRice Networks GroupA user can obtain an interference-free channel by sharing its channel information
Client 1Client 2User affected by APsinterferenceChannel InformationAPOne of the other benefits of ZFBF is that a user can obtain an int Connsider as an example an undesired user that is affected by tranmission from the AP.[animate] The undesired user may not even share the same standard as the rest of network, however, if it can somehow provide feedback on its channel, AP can serve its users while suppressing interference that goes towards the undesired userWe can explain how this works through our received signal channel formula OutlineBackground
System Implementation
Experimental Evaluation
Conclusion
Ehsan AryafarRice Networks GroupMethodologyUnified Implementation PlatformFirst Implementation and experimental evaluation of different beamforming algorithms on a common platform
Experimental Characterization of System PerformanceCompare against single-user TDMA schemesUse repeatable controlled channels andReal-time indoor channels
Evaluation MetricSNR or the corresponding Shannon capacity
Ehsan AryafarRice Networks GroupWARPLab Research FrameworkWARP is clean-slate MAC and PHYOff-the-shelf platforms: Limited programmability/observability WARPLab brings together WARP and MATLABManage network communication of up to 16 WARP nodes Baseband signals are generated in MATLAB and downloaded to WARP nodesWARP nodes send/receive the RF signals
Ehsan AryafarRice Networks Group
Virtex-II Pro FPGAWe have used the WARP platform for our implementation.WARP is developed by the graduate students at Rice. The most important component of the platform Is the vertex 2 processor. It has two power pc cores that provides an environment in which you develop MAC and higher layer protocols in C. The physical layer algorithms are also implemented in FPGA, making the system real-time. It also has for slots for dauther cards, allowing for 4*4 MIMO
This platform is especially suitable for cross layer design, where you can easily obtain any information among different layers. In addition, due to high programmability, you can easily implement different MAC/PHY algorithms.While WARP is suitable for implementation of different PHY algorithms, changing the framework for advanced physical lay ..WARPLab is a framework for rapid impl of advanced physical layer algorithms and brings together the benefits of WARP and MATLab.In this frameowk ImplementationEhsan AryafarRice Networks Group
TxRxRxRx
( )
BF Weights5MUBF Data (OTA)6Rx Training Feedback3Rx RSSI Readings7Log RSSI Data (End of Cycle)8H Matrix and Weight Calculation4Training1Training (OTA)2For more information about our testbed and implementation please attend our demo!Experimental DesignMultiplexing GainReceiver separation distanceUser selection algorithmUser population size
Channel VariationEnvironmental variationUser mobility
Spatial ReuseLocation based interferenceMulti-point interference reductionNetwork throughput
Ehsan AryafarRice Networks Group
Impact of Receiver SeparationEhsan AryafarRice Networks Group
Issue: How does receiver separation distance affect spatial multiplexing gain?
Impact of Receiver SeparationEhsan AryafarRice Networks Group
R2Issue: How does receiver separation distance affect spatial multiplexing gain?
R1Impact of Receiver SeparationIssue: How does receiver separation distance affect spatial multiplexing gain?
ZFBF doubles capacity compared to Omni
Similar capacity up to /2 Separation distance
ZFBF at /4:6 dB decrease in per-link SNR Ehsan AryafarRice Networks GroupLocation ID: 2 3 4() 5(/2) 6(/4) 7Location ID: 2 3 4() 5(/2) 6(/4) 7Experimental DesignMultiplexing GainReceiver separation distanceUser selection algorithmUser population size
Channel VariationEnvironmental variationUser mobility
Spatial ReuseLocation based interferenceMulti-point interference reductionNetwork throughput
Ehsan AryafarRice Networks Group
User MobilityEhsan AryafarRice Networks Group
Issue: Evaluate impact of outdated channel information due to user mobility
User MobilityIssue: Evaluate impact of outdated channel information due to user mobilityRepeatable channel conditions802.11n Task Group channel model
Required channel update rateChannel must be updated at (/8) movementEqual to 10 msec update rate for a typical pedestrian speed (3 mph)Ehsan AryafarRice Networks Group
Per-link SNRAggregate CapacitySNR (dB)bps/HzSimilar experiments can be done for static receivers (in paper). The required channel rate for a typical residential environment is 100 msec.Experimental DesignMultiplexing GainReceiver separation distanceUser selection algorithmUser population size
Channel VariationEnvironmental variationUser mobility
Spatial ReuseLocation based interferenceMulti-point interference reductionNetwork throughput
Ehsan AryafarRice Networks Group
Multi-Point Interference ReductionEhsan AryafarRice Networks Group
Issue: Evaluate a senders ability to reduce transmission footprint at multiple locationsInterference reduction at unintended receiversImpact on the QoS of the served user
p1Interference Reduction PointsMulti-Point Interference ReductionEhsan AryafarRice Networks GroupIssue: Evaluate a senders ability to reduce transmission footprint at multiple locations
Interference Reduction:Interference reduction capability does not depend on the location/number of unintended receivers
Multi-Point Interference ReductionEhsan AryafarRice Networks GroupIssue: Evaluate a senders ability to reduce transmission footprint at multiple locations
Interference Reduction:Interference reduction capability does not depend on the location/number of unintended receivers
Increase in number of unintended receivers, can significantly drop the QoS of the currently served users
SNR difference at the intended receiverPrior WorkTheoretical Work on MU-MIMODPC (Costa83) and its optimality (CS03)ZFBF (YG06 and WES08)
Practical ProtocolsIAC (GPK09) and SAM (TLFWZCV09) Ehsan AryafarRice Networks GroupWe present the design and experimental evaluation of a MUBF platform for wireless LANsThere is a rich body of work on MU-MIMOIn SummaryDesign and implementation of the first MUBF platform for WLANs and found via experimental evaluation:
Users can simultaneously receive data down to a half of wavelength from one another
ZFBF can tolerate channel variations due to environmental variation, however, is strongly affected by user mobility
ZFBF can efficiently eliminate interference at undesired locations. This does not depend on the location/number of unintended receivers, however, can significantly reduce the QoS for the currently served users
WARP: http://warp.rice.edu RNG: http://networks.rice.eduEhsan AryafarRice Networks GroupBack UpEhsan AryafarRice Networks GroupiburstEhsan AryafarRice Networks Group
Patented technology for concurrent transmissionSuitable for outdoor channelsCrash Course on BeamformingOmniFixed vs ant selectionEhsan AryafarRice Networks Group
p1p2APSwitched BeamFixed beamHigh coverageAdaptive BeamHigher rangeSUBF
Lets start on a quick background of how you can serve multiple people. In this figure we have an AP with two antennas and two clients each with one antennas.[animate] The traditional way of communicating with these users, is to serve one user at a time in a TDMA fasion.There are many different ways for the access point to use its multiple antennas. One of the simplest methods currently used by many Aps is use of a fixed antenna (or rando selection) by AP and transmit in an omni-directional way. Figure at the left shows the radiation pattern for this method. In the basic approach no feedback from clients is received.Off course, Aps could benefit from the available antenna diversity. There are several papers that have shown you can obtain huge gains by selecting the appropriate antenn. An alternative way for AP to benefit from its antennas is through switched beam. Here, the AP has an antenn array. The signal that is fed to each antenna is shifted with specific phases, such that the resulting beam has high directionality in a specific direction. The AP can then switch between a set of pre-defined beams. Compared to the Omni transmission, this scheme allows for higher received SNR at the receiver and also higher coverage.A user can further increase its received SNR by sharing its channel information to the AP and through adaptive beamforming scheme. If AP has a users channel information, it can construct its beam such that it points towards that specific user. This scheme can further increase received SNR and coverage with the additional feedback of channel information for each antenna.
Weight Selection AlgorithmsZero-Forcing beamforming (ZFBF)Condition: => Heterogeneous link qualities through power allocationRegularized Channel InversionIncrease system performanceDoes not easily allow for heterogeneous link qualities due to non-zero inter-user interference
Ehsan AryafarRice Networks Group
An alternative way to ZFBF is to allow some amount of inter-user interference by removing the zero-interference condition.This approach is termed reg .The benefit of this approach is that it increases the overall system performance.However, there is not a closed form solution for providing different link qualities through power allocation.The focus of this talk is on ZFBFThis is beacuase
Multi-Point Interference ReductionEhsan AryafarRice Networks GroupIssue: Evaluate a senders ability to reduce transmission footprint at multiple locations
Interference Reduction:SUBFs interference could be significantly higher/lower than Omni
ZFBFs interference reduction capability does not depend on the location/number of unintended receivers
Weight Selection Zero Forcing Beamforming (ZFBF)Assume 4 Tx Antennas and 3 single-antenna receivers
hk's H for each recv.Calculate weights with pseudo-inverse
wj'sZero Interference Condition
Single user Hconj/Norm(H)
Zero InterferenceThis is a mathematical representation of the zero-interference conditionensures that if I have weights for R1, the resulting gain of stream 1 at R2 and R3 is zero.
Now, lets move on to how we implemented this in hardware.30Implementation - WARPLab
All baseband processing performed on Host PC Processed signals are downloaded to buffers in FPGA on transmitting WARP node HostPC sends Transmit/Receive trigger signals to WARP nodes Data is transmitted over the air, stored in buffers on receiving nodes FPGA Data/RSSI readings uploaded to HostPC for data processing/loggingWARP/WARPLAB MODIFICATIONS
We implemented the boxed equation in the PHY itself because it allows us to skip re-downloading the beamformed data to the FPGA (a second long process) after channel estimation.
Avoids downloading 2^14 samples again, only have to download the weights, Ensures that the channel estimate is still valid31User Population Size
Aggregate CapacityAverage Per-User SINR Q: How does the number of concurrently served users affect performance?
A: Capacity increases and saturates while per-user SINR drops significantly. CMC Lab6 Nodes, 3 Tx/Rx, 3 just RxAll connected to Host PCAll combinations of transmitter and receiver sets are evaluated Round Robin styleWide variety of link qualities from varying relative distances, and different obstacles
From now on, all experiments will focus on two receivers. This is because two receivers is below the number of DOF in our system and in a real implementation, we can assume that a base station would have a large number of antennas 32User Selection (Link Quality Difference)
Q: How do link quality differences between receivers affect system performance?
A: Link quality differences between concurrently served users do not affect each users SINR.
Same CMC lab setupQuestion:This question may provide insight for scheduling algorithms. If I am going to concurrently serve two users out of a larger set, should I be sure to schedule high quality links together? Low quality links together? Mix and match?Graph:-As before, we take measurements round-robin style of all combinations of a transmitter and a pair of receivers (links).- For SUBF, each point is simply the SUBF SNR of a particular link- For ZFBF, each point represents the average SINR of a particular link when paired with all possible other links and is also graphed against its Omni link SNRZFBF is lower than SUBF because w/ SUBF, full transmit power is used to serve a single user w/ SUBF while that transmit power is split amongst the pair for ZFBF-On ZFBF the Red bars indicate the full range of that links SNR when paired with all possible other links.-The Green bars indicate the full range of all other paired link qualities (the Omni SNR)-We see that even through the SNR range for the paired link is very large for each main link, the range of the resulting SINR is negligible
33Environmental Variation
Aggregate CapacityAverage Per-User SINR802.11n Task Group model for indoor residential environment(T) : Typical Fading rate of 1.157 Hz(R) : Rapid Fading rate of 2.778 Hz Q: How does performance vary with channel update rate in typically/rapidly varying channels?
A: Assuming a link can suffer up to a 3dB decrease in SNR below Omni, 100ms and 50ms update rates are necessary for typically/rapidly varying channels, respectively. Describe channel estimate delay time
Very sensitive to Environmental variation
Rapidly changing environment, at omni CAP by 100 ms
Demonstrated that a maximum 10 ms channel update rate results in at most 3 db loss in SNR for individual links for receivers moving at pedestrian speed (3 mph)Demonstrated that a maximum 500 ms update rate results in at most 3 db loss in SNR for individual links in a typical indoor non-mobile environment.
Fading rate is doppler fading rate34 Q: How does MUBFs interference reduction capability vary with the location of the unintended receiver?
A: The location of the unintended receiver does not affect the interference reduction performance of MUBF (when #Rx < DOF).Interference Reduction (Location)
Interference at WTransmit to R, reduce interference at W
Similar to Recv Sep Dist 35Channel VariationEhsan AryafarRice Networks Group
TestbedEhsan AryafarRice Networks Group
Channel EstimationEhsan AryafarRice Networks Group
Network ThroughputEhsan AryafarRice Networks Group